How well can you hide your age? Computer scientists at the University of Illinois at Urbana-Champaign have developed a software program that estimates age based solely on someone’s facial appearance, suggesting that in the near future you won’t be able to fool either Mother Nature or that video camera verifying your ID at the local bar.
Beyond more accurate age estimates, the technology suggests a way for ads to target passersby with age-appropriate pitches, for face-based security systems to improve their accuracy, and for robots to become more adept at responding to human needs.
“Definitely for human-robot interactions, the robot would like to know as much about humans as possible,” said Thomas Huang, a professor of electrical and computer engineering at the University of Illinois who led the research.
Correctly guessing someone’s age is a huge challenge even for humans, a problem that can be complicated by cosmetic interventions, genetics, gender, ethnicity, drug or alcohol consumption, and even relative sun exposure. Computer algorithms based on appearance can only guess a person’s perceived age, but even rough estimates could be a boon for electronically managing customer relations or compiling demographic data.
Imagine, for instance, a strategic camera and age-recognition model calculating the percentage of McDonald’s Big Macs bought by males of a certain age group. Likewise, an interactive ad could be tailored to tempt older pedestrians with a McCafe hazelnut cappuccino instead of, say, a Happy Meal.
“I think we are moving closer and closer to ‘Minority Report,’ where we have these systems that look at you and can generate some demographic information automatically,” said Karl Ricanek Jr., director of the Face Aging Group at the University of North Carolina at Wilmington.
In the science fiction movie, shoppers are greeted by advertisements that know plenty about their personal tastes. Israel-based YCD Multimedia has already taken several steps toward turning fiction into fact with customized in-store ads based in part on each viewer’s perceived age.
Japanese company Fujitaka has taken a different tack by equipping some vending machines with video cameras and age-recognition software to prevent underage teens from buying cigarettes, though several news accounts suggest the system can be tricked by a picture of an older person taken from a magazine.
Huang said his software isn’t yet ready to tackle such a sensitive age-verification issue, and he conceded that even marketing applications may need more sophisticated programs. Identifying the right target audience among passing shoppers, after all, likely requires advertising software that can guess ages based on side profiles as well as frontal views.
Beyond busting underage drinkers or smokers and startling shoppers with eerily precise cosmetic or clothing ads, Ricanek said age prediction technology could be combined with facial recognition systems to improve border security.
Ricanek, whose work is independent of the University of Illinois research, proposed the following scenario: Suppose a terrorist had his passport photo taken when he was 30. Based on that photo, would surveillance cameras be able to recognize him when he crosses the border 8 years later? A method that takes aging into account just might.
Facing the signs of aging
Huang’s group approached the general aging problem with a “black box” approach that relied on computer algorithms to establish connections among people of the same or similar ages. For one study, an algorithm divided each face into 10,000 pixels and asked which pixels clustered among people of the same age group. Another algorithm looked for similarities and differences in the geometry of key facial features, and a third measured texture as a way to quantify the amount of age-related wrinkling.
In two separate papers, published earlier this year in the journals IEEE Transactions on Multimedia and IEEE Transactions on Image Processing, Huang and his colleagues trained and tested separate algorithm combinations primarily on a database of 1,600 individuals ranging in age from 1 to 93 (each person was represented by five pictures taken around the same time, for a total of 8,000 images).
The programs’ guesswork on new faces, while besting other available techniques, has been far from infallible. So far, Huang said, the software can accurately estimate age to within five years only half the time, though it correctly assigns ages to within a decade about 80 percent of the time. For one face shot of Albert Einstein taken when he was 33, the computer model was right on the money, while another estimate of his age when he was 55 was off by 15 years.
Huang said his group is refining the two algorithm approaches and hopes to improve their accuracy by training both on far larger databases, like the MORPH archive of 55,000 faces from 13,000 people maintained at the University of North Carolina at Wilmington.
Even the primary morphological mechanisms behind aging have been difficult to discern, said Ricanek, whose group manages the MORPH database. In children, the apparent aging process is mediated primarily by skeletal growth and development, while in adults, changes in soft, non-bony tissues predominate.
“If we lived in an ideal world — a perfect world — the only effects would be from our genes and gravity,” Huang said.
Effects of aging
But in the real world, aging can be easily affected by behavior as simple as raising your eyebrows. Over time, raising your brows gradually creates what are known to scientists as hyperdynamic facial lines in your forehead (wrinkles for the rest of us). Younger, more elastic skin can erase the creases, but after a while, older adults are stuck with the kinds of laugh lines, crow’s feet and yes, forehead folds that have sent celebrities running for their Botox injections.
Similarly, photoaging, or sun exposure, can begin to crack and crease someone’s skin as if it was a cowhide left out in the sun. Tanning while taking some prescription drugs can lead to even faster photoaging.
Ricanek is developing models to simulate the aging process and produce images of how someone might look 20 years later, though like Huang, he has based his research on estimating someone’s apparent rather than actual age. He figures his group’s methods can estimate age to within four or five years of the truth more than half the time, but he cautioned that his analysis cannot be directly compared with Huang’s methods.
As age-verification applications become more prevalent, privacy concerns are bound to spring up, though Ricanek and Huang both said computer software can address those concerns by capturing only data on gender and age and not the individual’s identity.
Consumers in the U.S. may still have some time to get used to the idea. For an age-identification system to be reliably used for something like detecting underage drinkers, Ricanek estimated it would need to have no more than one false alarm per every 100,000 tries — a success rate he believes will require a decade or so to achieve.